Effect of marginalizing flows on rational-quadratic neural spline flows

Abstract

Rational-quadratic neural spline flows offer higher flexibility for complex distributions than autoregressive flows with affine coupling layers. Having an effcient analytical onepass inverse, the flow has an advantage over masked autoregressive flows in terms of inverse efficiency. In this thesis, the rational-quadratic neural spline flow was implemented with the aim to extend the flow to a marginalizing flow and investigate the performance differences between the original and the extended flow. In this work, we show that the implementation was challenging and resulted in an inappropriately performing flow. Due to the lack in performance, the usage of marginalizing flows was obsolete at this point. In conclusion, the implementation will need to be further developed in order to establish marginalizing flows.